Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
ChangeLog: Features The new remove-empty-columns and skip-trivial-records are keystroke-savers for things which would other require DSL syntax, as tracked in #274. Bugfixes A bug regarding optional regex-pattern groups was fixed in #277. As of #294 you can now specify --implicit-csv-header for the join-file in mlr join. A bug with spaces in XTAB-file values was fixed on #296. A bug with missing final newline for XTAB-formatted files using MMAP files was fixed on #301. Documentation Look-and-feel at http://johnkerl.org/miller/doc/ is (hopefully) improved, including clearer visual indication of which section/page you're currently looking at. Note that this change has been live for a few weeks, as look-and-feel-related doc-mods from post-5.6.2 were backported to http://johnkerl.org/miller/doc/. #282 improves DSL-function documentation at http://johnkerl.org/miller/doc/reference-dsl.html#Built-in_functions_for_filter_and_put,_summary Note Support for mmap mode has been entirely discontinued. This is an invisible change and should not affect you at all. For anyone interested in lower-level details, though, the summary is as follows: For an incremental performance gain (perhaps 10-20% run time at most, but see below), within the C source code one can use the mmap system call to access input files via pointer arithmetic rather than malloc-and-memcopy using stdio. However mmap is not available when reading from standard input -- it cannot be memory-mapped. This means all file-format readers are implemented twice within the Miller source code. While I try to regression-test Miller thoroughly, running all canned tests through mmap and stdio mode, I've nonetheless found my mmap implementations liable to corner-cases which I miss but users find: for example #29, #102, and #296. As tracked on #160, various operating systems do not release mmapped pages after use as one might intuit, meaning that for large files and/or large numbers of files, I've for a long time now needed to have Miller opt out of mmap usage for precisely those cases which most need the performance gain: see #160, #181, and #256. Additionally, mmap is not used at all for Windows/MSYS2 so there is nothing to lose there. For these reasons, keeping mmap mode isn't worth the development overhead. As of release 5.6.3, the mlr executable will still accept the --mmap and --no-mmap command-line flags as no-ops, for backward compatibility. The caveat for you is that for everyday small files, the default was previously mmap mode and is now stdio (except mlr ... < filename or ... | mlr ... which have always used stdio). There is the off chance that this will newly reveal an old, latent bug or two somewhere. I've re-run regressions in valgrind mode to aggressively catch any errors, but, please let me know ASAP via GitHub issue of any unexpected behavior in 5.7.0.
- Loading branch information